Keep pulling the thread on Noufar Gaspar.
Meta employees used between 60 to 74 trillion tokens in a single month.
Uber launched an AI adoption leaderboard and consumed its entire 2026 AI coding budget in approximately four months.
An unnamed company incurred a $500 million cloud bill due to having no AI usage limits in place, according to TechCrunch.
McKinsey estimates that approximately 60% of an agentic task's cost is tied to checking, refining, and regenerating answers after the first response.
Anthropic's new tokenizer, released in April, produced roughly 30% more tokens for the same text according to the company's own documentation.
Independent analyses of over a million requests found that Anthropic's new tokenizer increased native token counts by 32% to 45%.
The change in Anthropic's tokenizer resulted in real-world bills growing by 12% to 27%.
In a Databricks experiment on coding tasks, the Sonnet model cost $2.09 per task, while the Opus model, which is more expensive per token, cost only $1.94 per task.
In a Databricks experiment, the Opus model was cheaper to operate than the Sonnet model because Sonnet required more iterations and reasoning to achieve the same results.
A Databricks experiment showed a more than 2x difference in cost per task with the same quality when running the same model through different agent harnesses.
The CFO of OpenAI proposed a scorecard metric called "useful intelligence per dollar".
Meta tracked employee AI usage on an internal leaderboard.